[Model] Move vision_feature_select_strategy into resolve_visual_encoder_outputs (#25938)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

vllm/model_executor/models/minimax_vl_01.py

@@ -17,7 +17,6 @@ from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.multimodal import MULTIMODAL_REGISTRY
 from vllm.multimodal.inputs import MultiModalFieldConfig
 from vllm.sequence import IntermediateTensors
-from vllm.utils.jsontree import json_map_leaves
 from vllm.utils.tensor_schema import TensorSchema, TensorShape
 
 from .clip import CLIPVisionModel
@@ -221,15 +220,6 @@ class MiniMaxVL01ForConditionalGeneration(nn.Module, SupportsMultiModal,
     def get_language_model(self) -> torch.nn.Module:
         return self.language_model
 
-    def _select_image_features(self, image_features: torch.Tensor, *,
-                               strategy: str) -> torch.Tensor:
-        if strategy == "default":
-            return image_features[:, 1:]
-        elif strategy == "full":
-            return image_features
-
-        raise ValueError(f"Unexpected select feature strategy: {strategy}")
-
     def _image_pixels_to_features(
         self,
         vision_tower: Union[CLIPVisionModel, SiglipVisionModel,
@@ -238,16 +228,10 @@ class MiniMaxVL01ForConditionalGeneration(nn.Module, SupportsMultiModal,
     ) -> Union[torch.Tensor, tuple[torch.Tensor, ...]]:
         # NOTE: we skip the step to select the vision feature layer since
         # this is already done inside the vision tower
-        image_features: tuple[torch.Tensor, ...] = \
-            tuple(vision_tower(p) for p in pixel_values)
-
-        def select_features(leaf: torch.Tensor):
-            return self._select_image_features(
-                leaf,
-                strategy=self.config.vision_feature_select_strategy,
-            )
-
-        return json_map_leaves(select_features, image_features)
+        feature_select_strategy = self.config.vision_feature_select_strategy
+        return tuple(
+            vision_tower(p, feature_select_strategy=feature_select_strategy)
+            for p in pixel_values)
 
     # adapted from https://huggingface.co/MiniMaxAI/MiniMax-VL-01/blob/main/modeling_minimax_vl_01.py#L616-L631
     def pack_image_features(self, image_features: list[torch.Tensor],